First of all self.w is shaped (n_features,) so the y_pred will be (n_samples,)
Since, y_pred shape is (n_samples,) and y shape is (n_samples,1) and when they subtract each other(y_pred-y) the dimension of the result is (n_samples,n_samples), whose dot product to X.T which is shaped (n_features,n_samples) gives result (n_features,n_samples) which was expected to be (n_features,).

Some places need to use iloc to slice instead of default accessors.

def _best_split(self, X, y, feat_idxs):
        best_gain = -1
        split_idx, split_threshold = None, None

        for feat_idx in feat_idxs:
            # iloc instead
            X_column = X.iloc[:, feat_idx]
            thresholds = np.unique(X_column)

            for thr in thresholds:
                # calculate the information gain
                gain = self._information_gain(y, X_column, thr)

                if gain > best_gain:
                    best_gain = gain
                    split_idx = feat_idx
                    split_threshold = thr
        return split_idx, split_threshold

Hi,
In the DecisionTree implementation, line 33 there is a call to the _most_common_label() method.
Line 33 is reached only if the y numpy array has one unique value, so instead of calling most common
you can simply take the value from any index of the array.

Maybe something like:

# check the stopping criteria
if (depth>=self.max_depth or n_labels==1 or n_samples<self.min_samples_split):
    return Node(value=y[0])

instead of :

# check the stopping criteria
if (depth>=self.max_depth or n_labels==1 or n_samples<self.min_samples_split):
    leaf_value = self._most_common_label(y)
    return Node(value=leaf_value)

Line 20 in PCA should change, eig returns eigenvalues , eigenvectors. It is stated in reverse form in the code.